misakid 0.1.0

Pure-Dart Misaki G2P engines with explicit resource and native adapter boundaries.

0.1.0

• Add immutable Kokoro handoff chunks that reproduce the pinned Kokoro 0.9.4 English and non-English text-splitting limits, fixed English unk='' policy, source-segment indices, and Python code-point behavior. An accepted eleven-case fixture executes the unmodified pinned pipeline while leaving vocabulary, voice/style, tensor construction, and ONNX inference to the consuming runtime.
• Add Android, iOS, and macOS bundled-native-asset build profiles to the Japanese Cutlet backend, with vendored checksum-pinned MeCab source, exact ABI/export validation, byte-backed ja_words.txt, and a repository-owned mobile release-build verification app. Android and unsigned iOS device artifacts are verified, including exact Android libc/libdl/libm linkage, and the complete accepted fixture passes through bundled assets on Android 15/API 35 arm64 and iOS 26.4 emulated runtimes. Physical-device and clean hosted-matrix validation remain before broad mobile release support is claimed.
• Add byte-backed English small-model and Chinese legacy/frontend-1.1 resource constructors so mobile applications can load packaged assets without filesystem-only APIs.
• Add the optional pure-Dart misakid_spacy_en adapter for the exact en_core_web_sm==3.8.0 tokenizer/tagger resources. It matches all 146 accepted English streams and 744 raw tokens without Python or native code.
• Add the optional misakid_espeak_en macOS-arm64 adapter for the exact eSpeak NG 1.52.0 tuple. Real spaCy-plus-eSpeak composition matches all 40 accepted fallback outputs, tokens, and 58 ordered raw calls.
• Add the experimental misakid_bart_en pure-Dart one-layer BART fallback adapter with explicit checksum-identified F32 resources, bounded parsing and inference, and a real Transformers 4.51.3 synthetic-model oracle. The named PeterReid weights remain unaccepted and unsupported.
• Accept the deterministic 33-member inventory for the exact 457 MB en_core_web_trf==3.8.0 artifact after checksum, archive-safety, metadata, and license review; the model remains an explicit external caller resource and is never bundled.
• Accept 116 American/British transformer cases across no-fallback and eSpeak fallback, with 2,278 captured raw token records, 50 ordered eSpeak calls, and exact final replay. Route the two resource closures through distinct locked oracle interpreters and add a deterministic graph/tensor inspection tool.
• Add the supported misakid_spacy_trf_en macOS 11+ arm64 adapter: exact tokenizer and byte-BPE stages in Dart, a package-owned Apple-Accelerate 12-layer transformer, strict external-resource identities, reproducible builds, and no Python or Torch runtime. The real backend matches all 116 accepted cases, and real transformer-plus-eSpeak composition matches all 40 fallback cases.
• Add exact transformer boundary cases at 104, 105, 144, 145, 208, and 249 marked pieces per dialect, including an emoji whose tag detects omitted overlap averaging.
• Bound transformer byte-BPE work by threading the configured aggregate piece limit through assembly and replacing quadratic full-list rescans with an exact ranked linked merge algorithm verified against the former reference scan on seeded adversarial tables.
• Correct both accepted eSpeak oracle locks to include upstream's exact joblib==1.4.2 transitive runtime dependency.
• Use generated CPython 3.12 lowercase semantics in shared English stages and add an authoritative Cherokee tag-weight regression whose final tag differs under the host Dart lowercase mapping.
• Complete both Japanese contracts on macOS arm64 with the explicit Open JTalk and MeCab/UniDic Cutlet adapters, including exact final/raw parity, resource identity, ownership, isolate, lifecycle, and no-stdio tests.
• Complete the pure-Dart Chinese frontend-1.1 backend without an English callback using eleven checksum-pinned resources and exact parity for all 26 accepted outputs plus every captured POS, pinyin, and search call.
• Support the exact Chinese frontend-1.1 plus pure-Dart English small-model no-fallback composition across American/British and legacy/2.0 rendering. Its 14 accepted cases exhaust 14 callbacks, 40 raw tokenizer records, and 33 final English tokens; the read-only aggregate now covers 431 cases in 17 fixtures through distinct combined and transformer/eSpeak interpreter keys.
• Exclude root /test and /tool content from the root library boundary so no Python oracle tooling, fixture tests, or fixture-generator inputs are distributed as runtime code.
• Keep workspace members directly publishable by allowing their publishable source trees through the root archive while each nested .pubignore removes tests, Python/oracle tooling, caches, models, and binaries. Required offline native-builder source and manifests remain in the relevant adapter archive.
• Add the optional misakid_chinese sibling package with a complete no-Python legacy/default backend: exact pure-Dart cn2an 0.5.23, Jieba 0.42.1 accurate+HMM, and pypinyin 0.53.0 stages over six explicitly supplied, checksum-pinned resources.
• Accept and replay 24 authoritative Chinese legacy fixtures, matching all 22 outputs, two pinned failures, 41 segmentation runs, 82 words/pinyin calls, 152 syllables, and null-token semantics; add a direct original-Jieba replay, adversarial resource/parser tests, complete MIT notices, and provisioned CI.
• Restore pinned g2pkc's non-overlapping silent-onset insertion for consecutive standalone Jamo vowels, fixing adjacent-ARPABET-vowel rendering and adding focused composition regressions.
• Make Vietnamese lower/upper and whitespace handling host-independent with exhaustive CPython-3.11.15/Unicode-14 candidate mappings, including multi-scalar expansion, supplementary scripts, and context-sensitive Final_Sigma; retain exact-3.11.13 comparison as an explicit parity gate.
• Restore pinned English decimal near-integer rounding, reverse-Japanese fractional unit/repeated-point quirks, and Cutlet's decimal-versus-digit Unicode distinction; harden injected English rating validation.
• Add 42 checksum-pinned adversarial English fixtures, bringing exact American/British no-fallback replay coverage to 106 cases across number flags and edge values, morphology, controls, Unicode, mixed scripts, and symbols.
• Add a strict accepted-fixture manifest and one read-only command that validates complete corpus/fixture/provenance coverage and verifies all 431 committed oracle cases by default.
• Harden the publish archive exclusions for lock, generated documentation, the empty bin directory, and Dart/Python cache artifacts; remove the misleading upstream Python homepage from package metadata pending a canonical Dart package repository URL.
• Pinned the exact en_core_web_trf==3.8.0 transformer oracle artifact and curated Python 3.12 dependency closure without downloading or bundling the 457 MB model, and prepared mode-only American/British transformer corpora.
• Recorded the checksum-pinned Mishkal 0.3.2 source/data license blocker; no historical Mishkal material is copied or translated.
• Recorded the GPLv3 underthesea 6.8.4 and independently unresolved packaged tokenizer-model boundary; no underthesea code, native library, or model is included.
• Add the optional misakid_openjtalk sibling package: an explicit no-Python macOS arm64 Open JTalk 1.11 adapter with verified source/dictionary identity, owned results, structured no-stdio failures, process-global locking, bounded inputs/results, deterministic close, and no committed native artifact.
• Pass its real production ABI against all 24 Japanese cases, 155 raw frontend records and 14 fields, all final phoneme/token fields, four concurrent isolates, direct native result ownership, lifecycle/bounds cases, and byte-empty native stdout/stderr; add reproducible release artifact gates.
• Add the optional misakid_mecab_ko sibling package: a reproducibly source-built macOS arm64 MeCab-ko adapter plus pure-Dart CMUdict provider, with exact source/data identities, owned no-stdio ABI, bounded lifecycle, and no committed native binary or data resource.
• Pass its production path against all 34 Korean morphology streams, 293 morphology records, 23 CMU lookups, 33 successful final outputs, and the pinned failure, plus concurrent-isolate, tampering, and lifecycle tests.
• Begin the behavior-preserving Dart port pinned to Misaki 0.9.4.
• Add shared engine, result, token, metadata, backend, and error contracts.
• Add dependency-free Japanese Arabic/kanji/kana/romaji number conversion.
• Add the pure Japanese pyopenjtalk-style mora, accent, pitch, and rendering pipeline behind an explicit frontend contract.
• Add the explicit Hebrew phonemizer backend contract and typed facade.
• Port the pure English phoneme stress and weighting stage.
• Port the dialect-sensitive English inflection suffix rules.
• Port exact English token merging with typed metadata.
• Port the right-to-left English vowel/to context stage.
• Port English subtoken resolution and stress balancing.
• Port exact legacy versus 2.0 English final rendering.
• Add the public injected English engine and typed tokenizer, pronunciation, and fallback backend contracts.
• Reject malformed English tokenizer records and any token stream that does not reconstruct the exact preprocessed input.
• Add the offline pinned American/British lexicon provider with contextual, morphology, currency, and clean-room number behavior.
• Expose immutable dialect/version English inventories and validate every generated gold and silver pronunciation against the pinned raw inventory.
• Pass both 52-case American and British no-fallback English suites end to end using the same strict typed raw-token replay boundary.
• Add pure eSpeak fallback replacement/postprocessing behind an injected raw eSpeak-phone backend without bundling or discovering eSpeak-ng.
• Pin and review the named US/GB BART fallback revisions and document the unresolved training-data, lineage, license, copyright, and notice blocker; no model artifacts are accepted or distributed.
• Add exact host-independent CPython 3.12.11 / Unicode 15.0.0 NFC/NFKC and alphabetic, decimal, digit, and whitespace tables with exhaustive scalar plus cross-scalar digest verification.
• Add the pure Japanese Cutlet normalizer and renderer behind a typed, explicitly grouped morphology boundary without vendoring ja_words.txt, with exact mapping integrity, typed-failure preservation, strict oracle capture scaffolding, and a compiling fixed-record example.
• Expose immutable source-derived Japanese phoneme inventories for both Cutlet and pyopenjtalk-style rendering, with exact scalar and mora-table digests.
• Port the pure legacy-Chinese tone and punctuation mapping helpers.
• Add public injected legacy and 1.1 Chinese engines with typed cn2an/jieba/POS/search/pypinyin boundaries, ordered sandhi and erhua, mixed-English handling, exact internal tokens, and null outer-token semantics. The legacy contract is now completed by the separate resource package; frontend 1.1 without an English callback is completed by the same resource package, and the precise small-model/no-fallback callback profile is completed by composing it with misakid_spacy_en.
• Expose immutable source-derived Chinese render-scalar inventories for the legacy and frontend-1.1 contracts, with exact digest and exclusion tests.
• Add the public injected Korean g2pkc engine with typed morphology and CMUdict boundaries, deterministic generated idiom/rule data, and null-token semantics.
• Pass the adversarial 34-case Korean default-mode matrix end to end using strict ordered MeCab and CMUdict replay: 33 exact successful outputs and the pinned 17-digit numeral failure boundary.
• Preserve CPython 3.12 seed-zero tuple-set iteration for order-sensitive Korean numeral replacements, backed by fixed hash/table vectors and a 4,096-case zero-mismatch offline differential.
• Expose the immutable 67-symbol modern conjoining-Jamo inventory used by the default Korean renderer and validate committed Jamo output against it.
• Add the public pure Vietnamese cleaner, phonology, substring behavior, clean-room number spelling, dialect/tone options, typed tokenizer and English-fallback boundaries, token metadata, and deterministic generated licensed data.
• Add strict raw eSpeak and Vietnamese external-stage oracle capture schemas for deterministic Dart replay.
• Add Linux, macOS, and Windows CI, warning-free API documentation checks, compiling public examples, and a versioned cold/warm latency, throughput, RSS, and generated-data-size benchmark.
• Enforce documentation for every public Dart member, including deterministic generated-data contracts, through fatal analyzer infos.
• Add a contract test that keeps the published library free of platform-only I/O, FFI, web, Flutter, and direct printing APIs.